Abstract

This paper discusses the utility of probabilistic representations for systems equipped with sensors and actuators. Probabilistic state estimation methods commit at no point in time to a single view of the world. Instead, they consider all possibilities, weighted by their plausibility given past sensor data and past actions. Probabilistic methods can represent uncertainty and handle ambiguities in a mathematically elegant and consistent way. In our own work, we have employed probabilistic representations combined with neural network learning for various state estimation problems that predominately arose in the context of mobile robotics. For example, probabilistic state estimation played a major role in our entry at the 1994 AAAI mobile robot competition [1, 18], where our robot "RHINO", shown in Figure 1, explored and mapped an unknown arena of approximate size 20 by 30 meter at a maximum speed of 90 cm per second. Probabilistic methods were also central for a recent exhibit, in which RHINO acted as a tour-guide for visitors of the Deutsches Museum in Bonn.

Citations